The Moving Line
Work lies on one timeline: what AI does, what AI can do but humans still do, what only humans can do, and the unbounded frontier beyond. Three boundaries move across it at diffusion, research, and frontier rates. This essay models who gets repriced, who evaporates, and who captures what opens next.
Introduction
Three claims about AI’s economic impact are circulating right now, each made by serious people, and they appear to contradict one another.
The first is a repricing claim, made by me from inside Nevo and, independently, by Anthropic’s Thariq Shihipar. Work that took a software consultancy three weeks — call it 120 billable hours — now takes six hours with frontier coding agents. That is a 95% reduction in time, a 20× speedup. If that ratio generalized across an implementation-heavy book of business, and clients continued paying by the hour, the result would be severe. Run the arithmetic on a delivery team of ten, billing at typical consulting rates at healthy utilization, and roughly $2.9M in annual revenue compresses toward $144K. On this view, firms must find twenty times the volume or watch their revenue collapse.
The second is an elasticity claim, stated most visibly by Jeff Bezos: AI will produce a labor shortage, not mass unemployment. The radiologists and software engineers everyone expects to disappear will instead be elevated — people who have been digging basements with shovels are about to be handed bulldozers. Productivity will rise so much that some households will choose to drop an earner, and core goods will get cheaper.
The third is a transformation-failure claim, made by Kai-Fu Lee: the overwhelming majority of enterprise “AI transformation” initiatives fail, because organizations bolt AI pilots onto functional tasks without ever touching the business core, and much of existing leadership is unequipped for the transition.
Here is what makes this interesting: I do not think any of these claims is wrong. I think each one is a correct observation of a different variable, and the variables move at different speeds. The apparent contradiction dissolves once you stop asking “which claim is true?” and start asking “which boundary and rate is each observer watching?”
This essay develops that conjecture into a model, and the model turns on a single timeline with four segments. First comes work AI does. Next comes work AI can do, but humans still do. Beyond that is work only humans can do. Beyond all currently defined work lies an unbounded frontier of problems, products, and roles still to be discovered. Three moving boundaries divide those four segments. The diffusion boundary separates actual AI use from latent AI capability. The research boundary separates what AI can do from what only humans can do. The frontier boundary separates defined human work from the unbounded possibility beyond it. The economic story is the story of their relative rates.
I am not a disinterested observer. I run delivery teams at an enterprise software consultancy, and this model describes the ground shifting under my own feet. That is a reason to scrutinize the argument harder, not a reason to discount it; the essay ends by asking what moves the model actually licenses for someone standing where I stand — and it will not conclude that anyone, including me, gets to stand on solid ground.
As always, this is conjecture offered for criticism. Every load-bearing claim below is stated as a comparison of rates, and each of the model’s predictions comes with a named condition that would refute it.
1. Rates, Not States
Most arguments about AI and jobs are arguments about states. “AI can write code, therefore programmers disappear.” “Demand for software is unlimited, therefore programmers thrive.” Both sentences compare a snapshot of capability to a snapshot of the labor market, and both are the wrong kind of claim, because the economy is not a snapshot. It is a system of processes running at different speeds, and what happens to any particular firm or worker depends on which processes reach them first.
So the method of this essay is to translate every static claim into a claim about rates. When someone says “AI does the work twenty times faster,” the useful questions are: can AI do it, does AI actually do it, and what new work is opening beyond it? This reframing sounds like a small move. It is not. It is the difference between predicting an equilibrium and predicting a path, and for anyone actually living through the transition — a firm, a worker, an industry — the path is the whole story. Equilibria do not send invoices.
Place all work on one timeline:
Work AI does | diffusion boundary | Work AI can do, but humans do | research boundary | Work only humans do | frontier boundary | Unbounded frontier
The segments are states; the boundaries carry the rates. Diffusion is the rate at which capable work becomes work AI actually performs in production. Research is the rate at which work moves from human-only into AI-capable. Frontier is the rate at which previously unbounded possibility becomes concrete human work: new problems, products, and roles that did not exist to be automated before.
The widths of the segments change as the boundaries move. When research outruns diffusion, the middle segment widens: AI can do more than organizations let it do. When diffusion catches research, latent capability becomes actual production and repricing follows. When the frontier opens faster than research advances, the domain of human work expands even while old work is compressed. A firm’s identity constrains how quickly it can move its own diffusion boundary, a mechanism developed in Section 7.
2. The Moving Line
The line most AI discussions notice is the research boundary, the line between work current AI systems can perform and work they cannot. Today, its left side has a recognizable profile: work that is specifiable, bounded, and verifiable, work you could hand to a competent stranger with a written brief and check without watching them do it. Its right side contains work entangled with organizational context, ambiguity, politics, trust, and accountability. Call the research boundary the compression frontier, because when it passes a kind of work, that work becomes technically compressible — the 120 hours becoming 6.
Nearly every confident claim about AI and work smuggles in the same hidden assumption: that this boundary is a fact about the work. It is not. It is a fact about the models — and models improve. Two years ago, writing production code sat comfortably in the human-only segment. Today it sits in the AI-capable segment, and in leading organizations it has crossed the diffusion boundary into work AI actually does. Verification — reviewing code, writing tests, checking invariants — was, until recently, the standard example of what would remain human; agents now write the tests and run the harnesses. The research boundary is not a property of tasks any more than a coastline is a property of the ocean floor. It is where the tide currently reaches, and the tide is coming in.
This means the method of Section 1 applies to the boundary itself. The wrong question about any kind of work is “is this compressible?” — a state question, and its answer expires. The right questions are “how fast is research making it compressible?” and “how far behind is diffusion?” Everything in this essay that looks like a static category — compressible versus serial work, depreciating versus durable advantage — should be read as a position on the timeline. Technical exposure is distance from the research boundary; realized disruption depends on the diffusion boundary reaching the work too.
A research boundary that moves raises the obvious question: does it ever stop? Is anything permanently in the human-only segment? I think the answer is yes — but the boundary of the permanent lies somewhere different from where most discussions place it. Distinguish work that is contingently beyond the research boundary from obligations that are constitutively human. Contingent work remains human-only because the models have not reached it yet — verification, as we just saw, was contingent all along, however permanent it looked. Constitutive obligations, by contrast, are not computations at all. They are social or legal relations. Liability is the clearest case: “someone must be accountable when the system fails” is not a task a better model performs but a fact about courts, insurance, and the human need to trust before depending. Even where law adapts to machine-made decisions, the accountability does not vanish — it relocates to whoever deploys, insures, or certifies the system. The form changes; the relation persists. Trust between institutions, the legitimacy of a consequential decision, the assignment of responsibility: a model can inform these relations, but it cannot be a party to them. Call this set of relations the constitutive core.
Two opposite errors haunt this distinction, and the essay will try to walk the ridge between them. Misclassifying contingent work as constitutive is the error that fills the graveyard of “AI will never X” predictions. Misclassifying constitutive work as contingent is the error that dissolves law, trust, and accountability into engineering problems they are not. How much of today’s uncompressed work is truly constitutive — most of the residual, or a sliver — is, I believe, the single most important open question in this subject, and I return to it in Appendix C. What matters for now is that every “non-compressible” claim in the sections ahead carries this distinction inside it: either the claim names a relation, or it names a position the tide has merely not yet reached.
With the timeline and its boundaries in view, we can now watch what happens when research makes work compressible faster than diffusion makes that compression real.
3. The Repricing Mechanism
Start with the napkin math from the introduction: a 120-hour task becomes a 6-hour task. (The arithmetic is laid out in Appendix A.) The naive reading says consulting revenue falls 95% unless volume rises 20×. The elasticity reading says volume will rise, because enterprises have roadmaps ten times longer than their budgets, and custom software may be the most demand-elastic good in the economy — every project that was uneconomical at $2.9M becomes thinkable at a fraction of that. Which reading is right?
Both — at different speeds. And here is the mechanism: potential supply responds to research, realized pricing responds to diffusion, and new volume responds to the frontier. Research can make a task cheap in principle, but price discovery requires only one competitor to move that task across its diffusion boundary and quote the AI-assisted price. Economy-wide volume arrives more slowly because most clients still occupy the middle segment: AI can do the work, but humans do it. Moving that work across the diffusion boundary requires change management, data readiness, compliance review, redesigned workflows, and organizational will — all of which run at firm speed, not model speed. An insurance carrier does not integrate twenty new systems this year merely because systems got cheap. And the eventual volume is not limited to today’s backlog; the frontier boundary opens problems and products that were not previously economical or even legible.
So the price falls fast and the volume arrives slowly, and the interval between them is what I will call the interregnum: the period in which the old revenue model is dead and the new demand has not yet materialized. The repricing claim correctly describes the interregnum. The elasticity claim correctly describes the equilibrium beyond it. Neither is a complete account, because each is watching a different boundary — a pattern we will see again.
The interregnum has a corollary that I think is underappreciated: the first casualty is not the consultancy but the pricing model. Time-and-materials billing works only while input time tracks output value. Once research and early diffusion decouple them, hourly billing becomes unstable regardless of what any individual firm decides — a firm that keeps selling hours into a market that has repriced hours is volunteering for the exposed class. The interesting question is therefore not “how do we do twenty times the volume?” but “how do we stop selling time?” And answering that question requires locating the firm’s work on the timeline — which brings us to Amdahl.
4. The Amdahl Inversion
Amdahl’s Law comes from parallel computing, and it says something simple: the total speedup of a system is bounded by the fraction of work that cannot be parallelized. If half your workload is inherently serial, infinite processors buy you at most a 2× speedup. Applied to consulting, the standard move is reassuring: coding compresses 20×, but discovery, stakeholder alignment, verification, and deployment do not, so the engagement as a whole speeds up far less than the headline number, and the revenue collapse is overstated.
The reassurance is valid when we hold one engagement fixed: its noncompressed work limits its total speedup. But a firm is not an engagement. A firm holds a portfolio of work, and the composition of that portfolio was chosen by a history of contracts and client relationships. The difference between those two levels is the inversion I want to develop.
Consider the firm level. The reassuring argument quietly assumes that every consultancy carries a healthy serial fraction of non-compressible work. But a firm’s mix of work is not handed down by nature; it is chosen, shaped by decades of contracts and client relationships. And in my own corner of the industry, the historical division of labor is stark: clients kept the serial work — planning, requirements, alignment, politics — and outsourced the specifiable core, the architecture and implementation. In many engagements the client-retained bottlenecks account for perhaps a tenth or less of total engagement effort. Why did the boundary land there? Because contracts govern best what can be specified, bounded, and verified — and so, over decades, the market naturally assigned consultancies the specifiable work.
Now recall the research boundary’s profile from Section 2: it advances first through work that is specifiable, bounded, and verifiable. The very properties that made implementation outsourceable make it automatable. This is not bad luck. The client–consultancy boundary was drawn, long ago, along precisely the research boundary’s current position — the firm’s territory contains the work research has now moved into the AI-capable segment. An implementation-pure consultancy is not partially exposed; it is maximally exposed by construction, because its portfolio was selected — by the logic of contracting itself — to contain almost nothing but the compressible fraction. At the firm level, Amdahl’s Law does not bind; a firm’s whole portfolio can sit between the research and diffusion boundaries, technically automatable but not yet fully automated in practice. The serial fraction is, in the economist’s sense, endogenous: a variable the firm’s history has chosen, not a constant imposed on it.
Now move inside the client organization. Within a particular transformation, someone must establish trust, satisfy regulators, change workflows, and hold accountability. No participant can choose those bottlenecks away at that moment, so the Amdahl analogy becomes useful again. This, I conjecture, is a deep reason diffusion is slow: its bottlenecks are exactly the organizational work of Section 1, and those bottlenecks run at organizational speed no matter how fast the research boundary moves.
But the economy is not one fixed transformation. Across organizations and across time, workflows are redesigned, tasks disappear, demand expands, and entirely new work appears. Amdahl helps explain why a given transformation remains slow without fixing the economy-wide quantity of work. It binds the path through a particular system; it does not conserve the system forever.
Put these levels together, carrying Section 2’s distinction, and a pattern emerges. As the research and diffusion boundaries pass contingent work, the remaining economic value tends to concentrate around constitutive obligations — trust, liability, and accountability. Call this the residual-concentration conjecture. The obligations persist, though the amount of labor and revenue attached to them need not.
The amount of serial work binding a transformation today is not the amount that will bind it tomorrow. The research boundary annexes the contingent portion generation by generation, and much of the work required for diffusion may prove contingent too — an uncomfortable possibility the model must carry rather than hide. What research never annexes is the obligation itself. That makes the constitutive core the residual’s likely destination without guaranteeing how economically large that residual will be.
The residual-concentration conjecture leaves two questions. How much economic value remains around the constitutive core? And how much of that value remains available for an outside firm to sell rather than being absorbed inside the client? A natural answer suggests itself immediately: the consultancy should simply move its portfolio farther right — stop selling implementation, start selling the residual. Hold that thought. It is correct as far as it goes, and Section 7 will show why it goes much less far than it appears. But first there is a second force reshaping the market, one that does not merely reprice work but removes it from the market altogether.
5. The Coasean Floor
Ronald Coase asked why firms exist at all — why any work happens inside organizations rather than being purchased, task by task, on the open market. His answer was transaction costs: using the market is not free. Finding a counterparty, negotiating terms, verifying quality, enforcing the agreement — all of this costs something, and when it costs more than doing the work in-house, the work stays in-house.
Every consulting engagement carries these costs in concrete form: procurement processes, master service agreements, security reviews, onboarding, access provisioning. Call their sum the engagement’s transaction-cost floor. The crucial property of this floor is that it is roughly fixed per engagement — a six-week procurement cycle costs the same whether it precedes six months of work or six hours of it.
Now run the compression from Section 3 against this floor. When the production work inside an engagement shrinks 20×, the fixed transaction costs come to dominate the total cost of engaging outside help. And past some threshold, the rational client does not respond by negotiating a lower price. The client stops contracting for that class of work entirely and does it in-house, because no price makes a six-week procurement cycle sensible for a six-hour task.
This is the mechanism I want to name clearly, because it is easy to mistake for repricing when it is something structurally different: the low end of the consulting market does not reprice — it evaporates. Work falls below the floor and exits the market. I am already watching the early form of this: clients running the same coding agents internally and asking, reasonably, why they need us. For the moment the answer is that we use the agents better than they do. The next section will examine how long that answer survives.
An existing relationship appears to offer an escape from this mechanism, and to a degree it does. An active master agreement and established trust lower the floor, so an incumbent vendor may delay evaporation by bundling compressed implementation work into work the client already trusts it to perform.
But the buffer extends only as far as the client’s understanding of what that relationship is for. It does not follow that a client who trusts a firm to implement against a specification will trust the same firm to redraw the specification, the workflow, or the organization around it. The identity problem of Section 7 appears inside the Coasean one. And where the client’s existing staff can wield the same capability, in-housing avoids both a new transaction cost and a revision of the vendor’s mandate. The existing relationship is therefore borrowed time, not an escape: it preserves work inside the old boundary without granting access to the new one.
Evaporation, not repricing, is why the right historical comparison for consulting is travel agencies rather than radiologists. Radiologists faced a capability threat and moved the capability across their diffusion boundary. Travel agencies faced a transaction-cost collapse — booking a flight yourself became easier than engaging an intermediary — and the mass-market product of routine booking was disintermediated, leaving a smaller market concentrated in corporate travel, complex itineraries, and high-touch advice. The consulting products that survive evaporation will share that profile: engagements too large, too risky, or too organizationally entangled to fall below the floor — transformation programs, regulated-system accountability, embedded capacity. Notice what these have in common: they cluster around the constitutive core from Section 2 — trust, liability, accountability — not artifacts. The floor independently rediscovers the research boundary’s residual.
One refinement matters for reconciling our three original claims. Repricing and evaporation strike the same firms, but the volume, when elasticity finally delivers it, arrives to different parties — and substantially to non-firms, meaning work that clients now perform internally and that never re-enters the external market at all. The Jevons-style recovery is real at the level of the economy and partially illusory at the level of the industry. This is how Bezos can be right about labor in aggregate while the consulting industry specifically has a very bad decade. Aggregate optimism and industry devastation are not in tension; they are the same event viewed from different distances.
How high the floor sits, and how it moves over time, is genuinely uncertain — research may compress transaction costs too, a possibility taken up in Appendix C. But whatever its height, a firm’s ability to stay above it depends on what kind of advantages the firm holds. The moving boundaries give that question a geometric answer.
6. Advantage as Distance from the Line
My firm’s current answer to “why do clients still need us?” is that we use coding agents better than our clients do. This is true today, and the timeline reveals exactly what kind of truth it is: the same work has crossed our diffusion boundary but not yet crossed the client’s. The advantage occupies the gap between two actors’ diffusion boundaries.
Each capability generation automates some of the technique previously supplied by an expert user. Early language models rewarded elaborate prompts; later models inferred more of the user’s intent for themselves. Prompt engineering was a genuine differentiator, but the models absorbed much of it within perhaps eighteen months. Orchestration skill — coordinating multiple agents, building verification harnesses — stands a little further up the beach, but on the same slope. The better the capability becomes at operating itself, the less durable an advantage based on operating it can be.
This suggests a classification every professional-services firm should run on its own balance sheet. The timeline supplies an ordering heuristic, not a literal calculation: an advantage’s half-life increases with its distance from the boundary that can erase it and decreases with that boundary’s velocity toward it. Tool-use skill of every kind — anything of the form “we are better at operating the capability” — is exposed from both sides. Diffusion teaches clients the technique, while research teaches the capability to operate itself.
What sits farther into the human-only segment? Only things whose distance is not measured in task difficulty alone. Proprietary context: knowledge of a client’s systems, data, politics, and history that is not reliably available to a general model, is expensive to transfer, and changes as the organization changes. Liability and trust: being the accountable party — constitutively human, in Section 2’s terms, and therefore outside a boundary that only crosses computations. Organizational position: being inside a client’s change process rather than a vendor to it. And distribution: owning the relationship through which capability flows to the client at all. Note the pattern: the durable advantages are not skills but relations — positions in a web of trust, knowledge, and access — which is exactly what the constitutive core predicted they would be.
Now the uncomfortable part. Look at where each class of asset accumulates by default. Tool skill accumulates in whoever uses the tools — for now, the consultancy. But context, trust, and position accumulate in whoever does the serial work — and Section 4 established that, under the historical division of labor, the serial work lives inside the client. The implementation-shaped consultancy has spent decades accumulating the asset closest to the moving boundaries while its clients accumulated the ones deeper in the human-only segment. This is the Amdahl inversion restated as an asset ledger, and it converts a strategy question into a solvency question: the firm’s remaining advantage has a half-life, the relative boundary rates are not under the firm’s control, and time is running.
The prescription seems to write itself: move up the ledger. Sell context, trust, and position instead of implementation. Move the firm’s work farther right on the timeline. I said at the end of Section 4 that this answer goes less far than it appears. Here is why.
7. The Double Bind: Value Networks and Identity
Clayton Christensen’s The Innovator’s Dilemma is usually summarized as “incumbents miss the new thing,” which misses his actual finding. Christensen showed that incumbents often see the disruption clearly and respond rationally — and that the rational response can still lead them into failure. A firm’s value network — its web of existing customers, pricing structures, sales motions, and margin expectations — evaluates every strategic option. Because the network rewards the business the firm already has, it reliably favors improvements that serve existing customers and preserve existing economics.
For a consultancy, the value network’s counsel is: use AI to deliver implementations faster at better margin. This is a genuinely rational response, and it is a melting-ice-cube strategy — improving margins on territory the research boundary has already made compressible. The alternative is to invade the client’s serial territory by selling discovery, diffusion, and organizational rearchitecting. But that means a different buyer, a different sales motion, and a different margin structure. The existing network votes against all three.
Whether AI-assisted consulting qualifies as disruption in Christensen’s strict low-end or new-market sense is not essential to the argument. What I am borrowing is his explanation of how a value network constrains an incumbent’s response even when the people inside the firm understand what is happening.
That is the first layer of the bind, and if it were the only layer, a sufficiently determined leadership could override it. The second layer is the one I have come to think is decisive, and it explains why diffusion rates differ so sharply between actors.
Call it identity — and I mean something specific, with an external and an internal face. Externally, identity is brand: the compressed prior in the client’s head about what your firm is for. A brand that has meant “hand them a spec, get back working software in a regulated environment” for twenty years cannot be re-indexed to “let them into our planning, our politics, our org chart” by a repositioned website. The client’s procurement function does not update its priors from marketing.
Internally, identity is culture — and culture is not the mission statement but the selection function: the accumulated pattern of who gets hired, promoted, and retained. A delivery culture selects, year after year, for people who excel at shipping against specifications. People who thrive in ambiguous, political, artifact-light diffusion work become rarer, and the traits they possess carry less institutional status. The firm that decides to sell diffusion may still employ some of those people, but it has not organized itself around finding them, empowering them, or reproducing what they do.
Identity constrains diffusion. A technical capability crosses a firm’s diffusion boundary only when the firm can operationalize it, sell around it, and reorganize the people accountable for it. A firm’s identity — what it sells, who it employs, what its name means — changes at a generational pace, because meaningfully changing a selection function means changing the population it has selected. Survival, for a firm, depends on moving its own diffusion boundary and its market position before research makes the old position worthless. The double bind is geometric: the value network votes against moving, and even a firm that overrules the vote discovers that its brand and people move more slowly than the research boundary.
Christensen’s own data supplies the empirical anchor, and it is grim. Incumbents rarely reposition the core organization into the disruptive segment. The successful exceptions generally create an autonomous organization — a protected business unit, joint venture, or spin-out — with separate economics and, tellingly, separate hiring. The case usually offered as the counterexample, IBM’s transformation into a services company under Gerstner, required a decade, a near-death experience, and an outsider CEO — which reads to me less like a refutation than like a statement of the price.
This yields what I will call, half-seriously, the insight-irrelevance theorem: in identity-bound transitions, insight is not the missing ingredient, and supplying more of it does not unblock the transition. The fix is not executable through the existing identity’s normal allocation mechanisms — not by better analysis, not by more conviction, not at any level of seniority — because identity constrains the firm’s diffusion rate, not the availability of knowledge.
This does not make leadership irrelevant. Leadership can create and protect a genuinely autonomous identity; what it cannot do is reason the existing organization into diffusing at research speed. I want to flag this as one of the boldest conjectures in the essay. Perhaps some mechanism can compress identity change without autonomy or near-death; Appendix C takes this up. But I have watched enough transformation initiatives to believe the strong tendency.
If the reported failure rate remains above 90% even as technical understanding spreads, identity offers an explanation the usual capability account cannot. The research boundary may be advancing while the organization’s diffusion boundary remains fixed. The claim that half of all companies need new leadership may therefore understate the problem. They need new identities, and leadership swaps are attempted precisely because identity swaps are rarely on the menu.
8. One Timeline, Three Boundary Rates
The complete model can be read from left to right:
Work AI does | diffusion | Work AI can do, but humans do | research | Work only humans do | frontier | Unbounded frontier
Each boundary converts the segment on its right into the segment on its left. Diffusion converts latent capability into actual production. Research converts human-only work into AI-capable work. The frontier converts unbounded possibility into defined human work. All three boundaries move, and the changing widths between them describe the transition.
Research currently moves fastest. It pushes specifiable, bounded, verifiable work into the AI-capable segment with each model generation. Diffusion follows unevenly. A new entrant can begin with its diffusion boundary close to research because it has no installed workflows or identity to preserve. An incumbent client may lag by years because capability has to pass through budgets, compliance, trust, and organizational change. The space between research and diffusion is therefore not a contradiction. It is a measurable stock of latent capability: work AI can do that humans continue to do.
The frontier moves in the same direction by making new human work concrete. A problem beyond the frontier cannot yet be assigned to either humans or machines because it has not been formulated as work. Once people turn it into a product, role, or tractable problem, it enters the human-only segment. Research may eventually reach it; diffusion may eventually operationalize it. But the frontier can keep opening new territory while both boundaries advance behind it. The labor question therefore depends on relative rates: whether research consumes defined human work faster than the frontier creates it, and whether diffusion realizes that technical possibility quickly enough to matter now.
The three public claims from the introduction each observe a different part of this motion. The repricing claim sees research advance and the earliest competitors diffuse it, enough to reset a market price. The transformation-failure claim sees the wide gap behind them, where capability exists but organizational diffusion stalls. The elasticity claim looks to the right edge, where falling costs move the frontier into problems and demand that were previously unreachable. Research explains what becomes possible, diffusion explains when it becomes real, and the frontier explains what becomes worth doing next.
The sequence of the interregnum follows from those relative rates. First, research moves the specifiable core of professional work into the AI-capable segment. Second, an early competitor diffuses the capability and collapses the price of hour-denominated work. Third, incumbent organizations lag in the middle segment, delaying volume and producing transformation failures. Fourth, where transaction costs remain high, the Coasean floor evaporates external engagements as clients move capable work across their own diffusion boundaries. Fifth, the frontier opens new demand, but to different parties and identities than the ones that were repriced. To the extent that constitutive obligations continue to command labor and economic value, they remain in the human-only segment even as the work surrounding them moves left.
A model assembled this neatly should make a reader suspicious, and the correct response to suspicion is to ask what the model forbids. Here is what it forbids.
9. Three Predictions
First: some agencies fail — and the model says which ones and how. The firms that fail first are implementation-pure agencies in weakly regulated verticals, priced in hours, holding no assets deep in the human-only segment — no context, no constitutive trust, no position. And the failure mode is specific: not the gradual margin compression a repricing story predicts, but discontinuous evaporation — whole classes of engagement vanishing from the pipeline as clients move capable work across their own diffusion boundaries and below the Coasean floor. Agencies in regulated verticals survive longer, held above the floor by trust and liability; but the model insists this is duration, not immunity, unless the borrowed time is converted into assets farther right on the timeline. This prediction is refuted if implementation-pure agencies broadly maintain revenue through the transition by volume expansion alone, without changing what they sell.
Second: new entrants win — for a time, and research sets the expiration. AI-native vendors, and in particular their forward-deployed engineering functions, capture the relocated residual because they arrive with no prior to overwrite: their brand already reads “we are the capability,” and their diffusion boundary begins close to the research boundary. They enter the diffusion market as greenfield players rather than as reformed consultancies fighting their own identity.
But “for a time” is not a hedge — it is a derivation from the boundary rates. Today’s AI-native diffusion product is tomorrow’s implementation shop, because much of diffusion work is contingently human: the research boundary that passed coding and then verification will press into deployment discipline, workflow redesign, and the translation of organizational intent into system behavior. The entrant’s founding advantage is a small gap between its diffusion boundary and the research boundary; only continuous movement keeps that gap small as its identity hardens into incumbency. This prediction is refuted if incumbent consultancies capture the majority of the diffusion market through their existing identities, without autonomous business units, joint ventures, spin-outs, or new companies.
Third: client-side workers have the strongest opportunity to capture upside — especially by moving. Look at where the assets in the human-only segment accumulate by default: proprietary context, organizational position, and trust all pool inside client organizations, in the people doing the serial work.
This does not mean the worker automatically captures their value. An employer can absorb the productivity gain through higher output expectations, lower headcount, or the codification of employee context into its systems. But the labor market gives the individual a redraw mechanism the firm does not possess.
Change employers and the worker arrives under a new prior: not as the incumbent employee expected to keep doing the old job more efficiently, but as the person hired to move the new organization’s diffusion boundary. The new role can reprice compensation, authority, and mindshare around that expectation, then place the worker inside the organization where context, trust, and position accumulate. This is identity-speed-of-one as labor-market arbitrage. The elevation thesis is most true, soonest, for client-side workers willing and able to use it. This prediction is refuted if workers hired explicitly to drive AI-enabled change fail to command greater compensation, scope, or organizational position while client-side technical roles stagnate or contract.
10. What Remains Executable
The model closes on a personal question, and I will not pretend it is hypothetical. If the insight-irrelevance theorem is right, then the person inside an incumbent consultancy who sees all of this clearly — the person whose job title says they are responsible for exactly this transition — cannot fix it through the existing unit’s normal allocation mechanisms. Not because they lack authority or conviction, but because the constraint was never insight. What moves does the model actually license?
I count two.
The first is to relocate on the timeline at identity-speed-of-one. Firms are selection functions with decades of momentum; individuals can change positions in months. A person can move into a vendor-side role close to the research boundary, or into a client-side role chartered to move the organization’s diffusion boundary. In either case, the new employer supplies a new prior — new mindshare, scope, and expectations — that reprices the individual around the new work rather than the old function. This reframes a question that deserves a principled answer: why would someone who sees the whole model leave rather than fix the firm? Because the fix is rarely executable through the existing identity’s normal mechanisms, while the labor market can reposition the individual immediately. Leaving is the third prediction applied reflexively, by someone arbitraging their own model.
The second is a new identity under a protected structure. This is the Christensen-compliant escape for a group rather than an individual: an autonomous business unit, joint venture, spin-out, or new company with separate economics and a brand carrying no implementation prior — the entrant advantage of the second prediction, claimed deliberately rather than stumbled into — selling the residual product from day one, in the regulated verticals where the Coasean floor sits highest and holds longest.
Honesty requires two caveats, and the essay’s own logic forces both.
The first concerns the second path. “Same people, new brand” satisfies the brand half of the identity condition and only partially satisfies the culture half — because culture travels with people, which is exactly why Christensen’s surviving spin-outs used separate hiring, not just separate letterhead. A new brand staffed entirely by implementation-selected people carries the old selection function inward even as it escapes the old prior outward. The model therefore predicts that the second path attempted as the old firm with a new logo fails by the same mechanism as the incumbents — and that it succeeds only as a genuine re-founding, in which the team’s composition is re-selected against the new product: diffusion work, ambiguity tolerance, political fluency, rather than delivery excellence alone.
The second caveat concerns the first path. If much of diffusion work is contingently rather than constitutively human — and I argued in the second prediction that it is — then research will, in time, press into forward-deployed engineering and client-side AI transformation roles too. The new role is not a safe harbor. It is a position: deeper in the human-only segment than implementation, with a longer half-life — but a half-life all the same. The honest statement of the model is that nobody in it stands on solid ground. There is no occupation, firm, or strategy in this essay whose durability is guaranteed; there are only positions on the timeline and rates at which its boundaries move. The individual’s advantage is not safety. It is the ability to reposition faster than a firm — and the willingness to use it more than once.
Which path is better? The Coasean floor locates the variables, but its height alone does not decide the answer. What matters is how the cost of using an outside firm changes relative to the client’s cost of moving the capability across its own diffusion boundary.
If production costs collapse while external transaction costs remain rigid, and the client’s staff can use the capability themselves, low-end work moves in-house. In that world, the first path points toward a client-side role chartered to lead diffusion. If transaction costs fall alongside production costs, small-scale external exchange survives, but the work becomes more fragmented and commoditized. That version of the first path points toward roles close to the research boundary or the frontier boundary rather than toward a firm selling hours. And if constitutive trust and liability keep some engagements large and valuable enough to remain above the floor, a protected new identity can sell the residual; that is the ground on which the second path can be built.
I do not yet know which of these forces dominates, and the answer will likely differ by vertical. The model does not choose the path for me. It tells me what to watch: internal diffusion costs, external transaction costs, and the amount of value that remains attached to constitutive trust.
Conclusion
I opened with three claims that appeared to contradict one another: consulting revenue collapses twentyfold; AI produces labor shortage and elevation, not unemployment; and the vast majority of AI transformations fail. The model developed here places all three on one timeline. Research moves work from human-only into AI-capable. Diffusion moves capable work into actual AI production. The frontier moves unbounded possibility into new human work. The repricing claim sees research and the earliest diffusion. The transformation-failure claim sees the gap where research has arrived but diffusion has not. The elasticity claim sees the frontier opening new demand beyond the work currently being compressed.
Along the way, the model produced what I think are its load-bearing ideas. The research boundary is a fact about models, not about work, and it moves — so every claim of “non-compressible” is a rate claim in disguise, while the constitutive obligations of trust, liability, and accountability persist rather than merely waiting to be annexed. The gap between research and diffusion explains how a capability can be economically decisive and organizationally absent at the same time. Amdahl’s Law exposes the bottlenecks inside engagements and organizations without fixing the quantity of work across the economy. Where external transaction costs remain high as clients move capable work across their own diffusion boundaries, the low end of the consulting market evaporates rather than reprices. An advantage lasts longer the farther it stands from the boundary that can erase it, and the durable advantages are relations, not skills. Identity constrains incumbent diffusion, offering an explanation for why failure persists even as technical understanding spreads, and why the residual may flow to entrants and insiders rather than to reformed incumbents.
If the model is right, its deepest irony is that the transition is hardest on precisely the organizations built to sell technical change, and offers the greatest upside to the client-side workers everyone expected it to displace — especially those willing to reprice themselves across an employer boundary. The bulldozer goes to the person standing in the basement. And its deepest discomfort is that the model exempts no one — not the entrants whose founding advantage ages from the day it is drawn, not the vendor-side engineer whose role merely sits farther right on the timeline, not the author. Nobody stands on solid ground.
I do not find this conclusion bleak, and I want to end by saying why, because the reason is not consolation — it is the same epistemology the essay runs on. The frontier boundary is generative. It turns previously unbounded possibility into problems, products, and roles that could not have been specified in advance — any more than “forward-deployed engineer” could have been specified five years ago, or “software consultant” fifty years before that. Research may move through each new segment in time, and diffusion may operationalize what research makes possible, but neither process bounds the frontier ahead. We navigate by conjecture: we can only pursue the opportunities we can currently see, knowing the timeline will extend and knowing that extension is precisely where new opportunities come from. The demand that our positions be permanent is a demand the world has never honored for anyone. What the world does offer — and what a fallibilist should ask of it — is the boundlessness of problems worth solving. Solid ground was never the offer. Motion was.
I hold all of this fallibly. The insight-irrelevance theorem is stated more strongly than the case law strictly proves; the size of the constitutive core is an open empirical question; the trajectory of the Coasean floor is another; and the graveyard of “the models will never do X” predictions counsels humility about every residual I have located, however carefully, in the human-only segment. The predictions in Section 9 come with their refutation conditions attached on purpose. If implementation-pure agencies sail through unchanged, if incumbents capture diffusion through their existing identities, or if boundary-facing client-side hires fail to capture greater compensation, scope, or position — the model is wrong, and I want to know it before the interregnum teaches the lesson the expensive way.
Appendix A: The Repricing Arithmetic
The compression figures used in the essay, for a representative engagement and delivery organization.
Task compression: original effort o = 120 hours (three weeks); new effort n = 6 hours. Remaining fraction r = n ÷ o = 0.05, i.e., 5% of original time, a 95% reduction, a compression ratio k = o ÷ n = 20.
Revenue exposure: a team of 10, at approximately 60% billable utilization across a 48-week year, at $250/hour, produces roughly 10 × 0.60 × 48 × 40 × $250 ≈ $2.9M annually. Under full 20× compression of the billable work at unchanged hourly pricing, the same delivered scope yields approximately $144K.
Two corrections from the essay apply to this arithmetic. First, the compression applies to the fraction of engagements in the AI-capable segment, not to engagements whole — though Section 4 argues that for implementation-pure firms that fraction approaches the entire book, making the naive figure closer to correct than the standard Amdahl objection suggests. Second, the arithmetic assumes hourly pricing survives the transition; Section 3 argues it does not, which makes the calculation a statement about the instability of the pricing model rather than a forecast of any particular firm’s revenue.
Appendix B: Key Terms
Work timeline. The ordered model: work AI does; work AI can do, but humans do; work only humans do; and the unbounded frontier. Three moving boundaries — diffusion, research, and frontier — divide the four segments.
Diffusion boundary. The line between work AI actually does and work AI can do but humans still do. Its rate measures how quickly technical capability becomes functioning production inside an organization or economy.
Research boundary (compression frontier). The line between work AI can do and work only humans can do. Its rate measures how quickly research makes human-only work technically compressible; its current profile favors work that is specifiable, bounded, and verifiable.
Frontier boundary. The line between defined human work and the unbounded frontier. Its rate measures how quickly previously unformulated possibility becomes concrete problems, products, and roles.
Unbounded frontier. The open-ended domain of problems and possibilities that have not yet become defined work. It is a segment beyond the frontier boundary, not a stock of existing tasks.
Contingently serial. Work that remains human-only because the research boundary has not reached it yet (e.g., verification until recently; likely much of diffusion work).
Constitutively serial. An obligation that persists because it is a social or legal relation rather than a computation: liability, accountability, trust between institutions, the legitimacy of consequential decisions. The constitutive core (Section 2) is the set of such relations. Their persistence does not guarantee how much human labor or economic value will remain attached to them.
Identity. A firm’s brand (the compressed prior in the client’s mind about what the firm is for) and culture (the selection function determining who is hired, promoted, and retained). Identity constrains the firm’s diffusion rate and its ability to reposition what it sells as research advances.
Identity-speed-of-one. The ability of an individual to reposition through an employer change in months while a firm’s selection function takes years to change. A new employer applies a new prior, repricing the person around the boundary-facing role they were hired to perform rather than the function they historically performed.
Interregnum. The interval between fast price compression and slow volume arrival, during which the old revenue model is dead and the new demand has not yet materialized.
Residual-concentration conjecture. As contingent work is compressed, the remaining economic value tends to concentrate around constitutive obligations — trust, liability, and accountability. The obligations persist, but the labor and revenue attached to them are not conserved. The serial fraction is endogenous (chosen) at the firm level and exogenous (binding within a given transformation at a given moment), while the economy-wide workload continues to change.
Coasean floor. The relationship-specific transaction costs of external engagement — procurement, contracting, security review, onboarding — below which work exits the market entirely (evaporation) rather than repricing. An established relationship lowers the floor within its trusted mandate; the vendor’s identity limits how far that mandate can be redrawn.
Advantage half-life. An ordering heuristic: an advantage lasts longer the farther it stands from the boundary that can erase it and the slower that boundary moves toward it. Tool skill is exposed to diffusion from one side and research from the other; the durable advantages — proprietary context, liability and trust, organizational position, distribution — are relations rather than skills, located deeper in the human-only segment.
Insight-irrelevance theorem. In identity-bound transitions, insight is not the missing ingredient; the transition is not executable through the existing identity’s normal allocation mechanisms, regardless of anyone’s level of understanding. Leadership can still create and protect a genuinely autonomous identity.
Appendix C: Open Problems
The size of the constitutive core. How much of the human-only segment — diffusion, trust work, accountability, judgment — is constitutively human rather than merely unreached by research? This is the question on which every durability claim in the model depends. Misclassifying contingent as constitutive repeats the “AI will never X” error; misclassifying constitutive as contingent treats law, trust, and institutional legitimacy as engineering problems, which they are not. The strongest candidate for a genuinely constitutive skill (as opposed to relation) is taste under uncertainty — judgment about which conjectures are worth the compute — and I hold it loosely, because it may only be contingent work farther to the right on the timeline.
The trajectory of the Coasean floor. The model’s market-structure predictions — and the choice between its two executable paths — hinge on whether research compresses transaction costs (procurement, verification, inter-firm trust) as quickly as it compresses production costs. An established relationship lowers the floor for work already inside its mandate, but identity governs whether that trust extends to repositioned work. If transaction costs and identity priors fall in tandem, small-scale market exchange survives and implementation work gig-ifies rather than in-housing. If trust and liability hold transaction costs rigid while existing staff move capability across the client’s diffusion boundary, the in-housing wave dominates. The answer likely differs by industry, with regulated verticals holding the floor highest longest. Note that this question is the constitutive-core question in disguise: transaction costs are rigid exactly insofar as they are constitutive.
The research boundary’s velocity and shape. The model treats research as advancing along a single gradient (specifiable → ambiguous). Is that right, or does it advance unevenly — leaping past work assumed safe while stalling on work assumed doomed? An uneven research boundary would complicate even the qualitative half-life ordering and demand a richer geometry.
The relative boundary rates. Does research move through defined human work faster than the frontier turns unbounded possibility into new human work? How quickly does competitive pressure help diffusion close the gap behind research? The answers determine whether the human-only segment expands or contracts, how long the interregnum lasts, and how much of the eventual demand reaches the parties that were originally repriced. The electrification precedent (roughly forty years from clearly superior capability to realized productivity, because factories were architected around the steam shaft) suggests a long diffusion lag; modern competitive feedback is the main reason to expect it to close faster.
Accelerating identity-bound diffusion. Autonomous units create a new identity alongside the old one; can the core organization move its diffusion boundary faster — deliberately, without autonomy, near-death, or outsider leadership as in the IBM case? If so, the insight-irrelevance theorem weakens further, and some incumbents can escape without re-founding. I know of no clean example, which is evidence, but not proof, for the stronger version.